Skip to content

README.md

Latest commit

 

History

History
47 lines (38 loc) · 1.91 KB

README.md

File metadata and controls

47 lines (38 loc) · 1.91 KB

PyroGuardian — Edge-AI Computer Vision System

Overview

PyroGuardian is an advanced wildfire detection system designed for real-time inference on edge devices (NVIDIA Jetson Nano). It utilizes state-of-the-art transformer-based object detection (RT-DETR) and is optimized via NVIDIA's DeepStream SDK and TensorRT.

Architecture

The system consists of two primary implementations:

  1. Baseline (YOLOv5): A PyTorch-based implementation used for research benchmarks and comparison.
  2. Advanced (RT-DETR + DeepStream): The production-grade engine optimized for Jetson Nano, featuring GStreamer pipelines and FP16 quantization.

Project Structure

  • deepstream_app/: Core C++/Python GStreamer pipeline for Jetson deployment.
  • tao_training/: NVIDIA TAO Toolkit configurations for training the 86.7M parameter RT-DETR model.
  • utils/: Shared utilities including the AWS SNS notification manager.
  • baseline_yolov5/: Original YOLOv5 implementation (Baseline).
  • dataset_tools/: Scripts for curated dataset augmentations.

Performance

  • Inference Speed: 30 FPS at 720p resolution on Jetson Nano.
  • Optimization: 90% speed-up achieved via TensorRT FP16 quantization compared to PyTorch FP32.
  • Accuracy: Robust detection across 8 fire conditions (Smoke, Forest, Urban, etc.).

Setup & Usage

1. Running the Optimized DeepStream Pipeline (Jetson Only)

cd deepstream_app
python3 fire_detection_pipeline.py <YOUR_AWS_SNS_ARN>

2. Training via NVIDIA TAO

# Requires TAO Toolkit Container
tao rtdetr train -e tao_training/rtdetr_train_spec.yaml -k $API_KEY

3. Running the Baseline

cd baseline_yolov5
python main.py --weights yolov5s.pt --source 0

Tech Stack

  • Models: RT-DETR (ResNet101 Backbone), YOLOv5.
  • Optimization: TensorRT, CUDA, FP16.
  • Deployment: DeepStream SDK, GStreamer, NVIDIA Jetson Nano.
  • Cloud: AWS SNS (Simple Notification Service).